Full support for AMD's out of core on GPUs
currently in Cycles, unless I've understood incorrectly (which I might have as I don't have an AMD gpu to test), only nvidia graphics cards reliably support out of core rendering (use system ram when vram runs out so you can still render entirely on gpu).
The next gen AMD gpu's are set to destroy the 3080 ti's according to most tech channels, but the problem is that without out of core support in Cycles, buying AMD over Nvidia would be a mistake. It's a shame, because AMD will most likely give blender users the opportunity to have much greater power for a much lower price than Nvidia.
It's not an opencl limitation as Radeon Prorender supports out of core with no issues, so cast your vote if you'd like the opportunity to potentially get more power for less money without limiting the size of scene you can render on the GPU with Cycles.
2 Comments
As far as I know, right now it is only supported by CUDA, see /rBc621832 , so it not even all Cycles's supported Nvidia GPUs, and it seems quite experimental and not advantageous in every cases anyway considering the huge performance loss once rendering using the system memory.
It's not like there is a will to allow only CUDA devices to render out of core, I haven't seen any trace of a dev pretending they wouldn't want it. It's just what's been implemented yet, like often on Blender, CUDA is first.
As to why this isn't on OpenCL and Optix yet, there could be several reasons, like it's harder/longer to do, the gains aren't considered worth putting the time and effort yet, .... Or maybe it's just planned, but way down the to-do list. There are quite some other things to consider first.
Like Nvidia GPU rendering no longer supported on macOS thanks to Apple's dropped support for CUDA drivers in macOS 10.14.
Or enhancing OpenCL rendering performances which are way behind CUDA and OptiX (which I think is far more important to deal before even thinking about OOC rendering).
Besides, I'd like to know how the "tech channels" can predict how some not-yet-released hardware will perform in a specific task, considering that it's never just a matter of values in a specification table.
And AFAIK GPU rendering often has been faster on CUDA rather than OpenCL devices because of rendering engines often being developped arround CUDA first.
* @-L0Lock- * yep it's a game changer on Cuda. I've been seeing 400% faster renders on my 1070 when out of core vs my i7-7700k, and since blender 2.79b I've been able to render over 30gb scenes (reported in render window) without a hiccup...in both gpu and hybrid modes!
Cycles is a beast.
Redshift has out of core on optix now, but not for geometry.